一种基于检错音网络的发音错误检测新算法  

An Application of Mispronunciation Detecting Network for Computer Assisted Language Learning System

作　　者：安丽丽[1] 吴延年 刘志 刘润生[3] 

机构地区：[1]清华大学微电子与纳电子学系,北京100084 [2]北京凌声芯语音科技有限公司,北京100085 [3]清华大学电子工程系,北京100084

出　　处：《电子与信息学报》2012年第9期2085-2090,共6页Journal of Electronics & Information Technology

基　　金：国家863计划重点项目(2008AA010700)资助课题

摘　　要：该文通过将计算机辅助语言学习(Computer Assisted Language Learning,CALL)系统的标准发音网络加入插入和删除路径的方法提出了一种发音错误检测新算法:检错音网络(Error-Detecting Network of Pronunciation,EDNP)错误检测算法。该算法首先对待测语音进行EDNP错误检测,然后通过对检错音网络的二级识别结果使用一级多候选词图进行错误召回的策略,进行错误检测,该算法易于实现并具有平台无关性。实验结果表明:该文提出的EDNP错误检测方法在中国四级考生语音测试库上使删除错误虚警率和漏报率分别达到7.38%和12.25%,插入错误虚警率和漏报率分别达到4.94%和26.17%,且客观评分与专家评分相关度比强制对齐方法的相关度提高了4.29%。This paper put forth a technique called Error-Detecting Network of Pronunciation （EDNP） that is applied to Computer Assisted Language Learning （CALL） system. By comparison with state-of-the-art CALL systems, the application of this kind of network is to insert mispronunciation detection routes into task-specific Finite State Grammar （FSG） network and avoid constructing complex mispronounced models. The detailed procedures of how to construct mispronunciation detection network and how to perform an error callback strategy are introduced in this paper. The algorithm is simply to be implemented and is independent to any speech toolkit. The experiments show that the application of this network achieves a False Acceptance Rate （FAR） of 7.38%, as well as a False Rejection Rate （FRR） of 12.25% for the deletion errors and achieves a FAR of 4.94%, as well as a FRR of 26.17% for the insertion errors. Furthermore, compared to traditionM forced alignment, there is 4.29% improvement to correlation rate between the objective and the subjective pronunciation quality evaluation by using EDNP.

关 键 词：检错音网络 发音错误检测 计算机辅助语言学习 

分 类 号：TN912.34[电子电信—通信与信息系统]